Covariance and contravariance in generic types
Static typing is awesome. It helps to detect bugs, acts as in-code documentation and makes development more enjoyable. Recently I've started to use Python's typing module to add static typing to all of my Python projects. Python's typing system may not be as as powerful as one might hope, but I think once you go typed, you don't go back.
It is, however, easy to run into non-intuitive errors when defining types for generic types such as lists and dictionaries. Assume, for example, that a
DogContainer takes a List of
Dogs in the constructor:
Now let's add some corgis to the container:
Type-checker raises the following error:
Argument of type 'List[Corgi]' cannot be assigned to parameter 'dogs' of type 'List[Dog]': 'Corgi' is incompatible with 'Dog'.
Why doesn't this work? Surely one can substitute a list of corgis for a list of dogs?
Let's step back and ponder why we think we can substitute
Corgi is a subtype of
Dog (we write
Corgi <: Dog, meaning we can use an instance of
Corgi wherever an instance of
Dog is expected), we think it should follow that
List[Corgi] <: List[Dog]. This property is called covariance.
C[T]is covariant in type variable
Tif the following holds:
A <: B ⇒ C[A] <: C[B]
We saw above that
List[Corgi] is not a subtype of
List is not covariant in its type variable.
List not covariant? The problem is that one can add new objects to
List, i.e., it's mutable. Or to put it another way,
List is not covariant because it acts as a sink of its type variable. Think of the following code:
class StBernard: ... def do_stuff(dogs: List[Dog]) -> None: """Function that secretly adds Beethoven to the input list. """ dogs.append(StBernard(name="Beethoven")) corgis: List[Corgi] = ... do_stuff(corgis) last_corgi = corgis.pop() # BOOM, type-safety broken! You think you have a corgi but you have Beethoven
The problem here is the
append method. What if we removed any methods from
List that make it mutable? That would essentially give us a
Sequence, an immutable list.
Sequence is different from
List in that it's a pure source of data. Intuitively, it is clear that any piece of code expecting a source of dogs would be happy to receive a source of corgis instead. Therefore, we can state that
A <: B⇒
Sequence[A] <: Sequence[B]
Sequence[T] is covariant in
T. Another example of an immutable container is
Mapping, a read-only version of
Pure sources are covariant
Let us generalize the previous discussion to the following rule of thumb:
If a generic class
C[T]is a pure source of instances of type
T, it is covariant in
Why are sources covariant? Assume you have
A <: B. Now imagine you have any code that depend on
Source[B]. Can you substitute
Source[A]? Yes, because
Source[A] produces instances of type
A that are valid instances of
For example, think of a class
Breeder[T] that produces dogs of type
T. Covariance means that
Breeder[Corgi] <: Breeder[Dog]. In other words, any code expecting a
Breeder[Dog] would be happy to receive a
Breeder[Corgi]. This makes sense intuitively.
The following piece of code shows how we could define a
Breeder class and mark it covariant in its type variable with
import abc from typing import TypeVar DogType_co = TypeVar('T', bound=Dog, covariant=True) class Breeder(abc.ABC, Generic[DogType_co]): @abc.abstractmethod def get(self) -> DogType_co: ... class CorgiBreeder(Breeder[Corgi]): def get(self) -> Corgi: print("Breeding a corgi") return Corgi() breeder: Breeder[Dog] = CorgiBreeder() # cast allowed because of covariance
If you look at the code from above where type-safety was broken, you may notice that another way to achieve type-safety would be to prevent reading from
corgis. In other words,
List should be write-only or a sink.
As a concrete example of a sink, consider the
Photographer class where the generic type must be a
from typing import Generic,TypeVar DogType = TypeVar('DogType', bound=Dog) class Photographer(Generic[DogType_contra]): def photograph(self, dog: DogType_co): print("Photographing dog") class CorgiPhotographer(Photographer[Corgi]): def photograph(self, corgi: Corgi): self.feed_sausage_to(corgi) super().photograph(corgi) def feed_sausage_to(self, corgi: Corgi): print("Feeding sausage to corgi") class DogPhotographer(Photographer[Dog]): def photograph(self, dog: Dog): super().photograph(dog)
Can we assume
Photographer[Corgi] <: Photographer[Dog]? This would mean that anywhere we need a dog photographer, we can use a corgi photographer. This is not the case: a corgi photographer cannot photograph a St. Bernard.
However, the inverse does hold: If one needs a photographer of corgis, a photographer of generic dogs definitely can handle the job. Formally,
Photographer[Dog] <: Photographer[Corgi].
Photographer is an example of a class that is contravariant in its type variable.The formal definition is:
C[T]is contravariant in type variable
Tif the following holds:
A <: B => C[A] :> C[B]
The way to mark the class to be contravariant in its type variable is to set
Thanks to contravariance, we can perform casts as follows:
Pure sinks are covariant
Photographer is contravariant in its type variable because it's a pure sink for dogs. It does not produce new dogs, but only consumes them (eww). We arrive at the following rule of thumb:
If a generic class
C[T]is a pure sink of instances of type
T, it is contravariant in
Why are sinks covariant? Assume you have
A <: B and imagine you have any code that depends on
Sink[A]. Can you substitute a
Sink[B]? Yes, because
Sink[B] is prepared to consume any
B and, because
A is a subtype of
B, it can also consume instances of type
Generic types that are neither covariant nor contravariant are invariant in their type variables. We saw above that
List is an example of an invariant generic type.
As another example of an invariant generic type, consider a
RescueShelter[D] class. An instance of
RescueShelter[D] accepts new poorly treated dogs of type
D and delivers them to their new homes.
RescueShelter is both a source and a sink of dogs, so it's neither contravariant nor covariant. One cannot substitute
RescueShelter[Corgi] does not accept dogs other than corgis. Similarly, one cannot substitute
RescueShelter[Corgi], because a person looking for a corgi from rescue shelter would not be happy to get a St Bernard dog.
This example illustrates the danger of relying too much on intuition. Intuitively, it might seem clear that a rescue shelter of corgis "is" a rescue shelter of dogs, and one might substitute the former for the latter. We saw above that this is only true if the rescue shelter only acts as a source of dogs.
In conclusion, it's generally easier to work with generic types that are either covariant or contravariant. Next time you define a function or data structure depending on
Dict, think carefully if you need to mutate the object. If not, you should probably use
Mapping instead. Similarly, if you do not need to read from your generic type, you may want to define it as write-only and mark the type variable as contravariant.
Thanks for reading! If there's anything here that seems wrong or inaccurate, please leave a comment.
Remember to enjoy corgis responsibly.